Listen to the article
Yandex has introduced a generative artificial intelligence model named Sona, designed to handle content recommendations by consolidating tasks that previously required dozens of separate algorithms into a single system. The company announced the model on 3 October 2026.
In conventional recommendation architectures, distinct algorithms divide the work: some sift through millions of candidates to find relevant options, others score those candidates, and still others determine the order in which they are presented to the user. Sona performs candidate selection and ranking within one unified model, eliminating the need to coordinate between separate components.
The model was trained from scratch on Yandex’s own data. Engineers from Alice, Yandex Music and the company’s research and development team all contributed to its creation and testing.
According to Yandex, one practical advantage of the single-model approach is that it removes the need to manually select the features the system relies on when forming recommendations. Sona analyses what a user has previously chosen and which content they have returned to, then uses that information to generate personalised suggestions.
One week on Alice speakers
Engineers tested Sona specifically for music recommendations on Alice-enabled smart speakers, replacing the standard multi-algorithm system with the new model for one week in a test group. Yandex reported that listening time for recommended tracks rose by 6.3 percent during the trial period, while requests to replay songs increased by nearly 18 percent.
The company also noted that consolidating recommendation logic into a single model simplifies future updates. Under the traditional approach, individual algorithms must be retrained and then re-synchronised with one another; Sona brings those tasks under one roof.